MATHEMATICS OF COMPUTATION Volume 66, Number 220, October 1997, Pages 1461{1486 S 0025-5718(97)00902-2 NUMERICAL SOLUTION OF ISOSPECTRAL FLOWS MARI PAZ CALVO, ARIEH ISERLES, AND ANTONELLA ZANNA Abstract. In this paper we are concerned with the problem of solving nu- merically isospectral flows. These flows are characterized by the differential equation L0 =[B(L);L];L(0) = L0; where L0 is a d d symmetric matrix, B(L) is a skew-symmetric matrix × function of L and [B; L] is the Lie bracket operator. We show that standard Runge–Kutta schemes fail in recovering the main qualitative feature of these flows, that is isospectrality, since they cannot re- cover arbitrary cubic conservation laws. This failure motivates us to introduce an alternative approach and establish a framework for generation of isospectral methods of arbitrarily high order. 1. Background and notation 1.1. Introduction. The interest in solving isospectral flows is motivated by their relevance in a wide range of applications, from molecular dynamics to micromag- netics to linear algebra. The general form of an isospectral flow is the differential equation (1) L0 =[B(L);L];L(0) = L0; where L0 is a given d d symmetric matrix, B(L) is a skew-symmetric matrix function of L and [B(L)×;L]=B(L)L LB(L) is the commutator of B(L)andL. The choice of the matrix function −B(L) characterizes the dynamics of the un- derlying flow L(t). Important special cases are the Toda lattice equations, double- bracket flows and KvM flows. Toda lattice equations in the Lax formulation (1) were considered by Toda [T], Flaschka [F] and Moser [Mo] and their relation with the QR algorithm for finding eigenvalues by Symes [Sy] and then extensively by Deift, Nanda, Tomei et al., Lagarias, in [Na1], [Na2] [DNT], [L], [DRTW]. It has been finally generalized to the nonsymmetric case by Chu, Watkins and Elsner in [Ch], [W], [WE]. The double bracket flow was introduced by Brockett in [B1] and then investigated by Brockett et al. in [BBR]. Its relation with the singular value decomposition (SVD) was considered by Chu, Driessel, Moore, Mahony, Helmke, Watkins, and others (cf. [ChD1], [D], [HM], [MMH], [WE]). Driessel and Chu in [ChD2] have also investigated another isospectral flow of the form (1) in relation with the inverse eigenvalue problem for Toeplitz symmetric matrices. Finally we mention the KvM Received by the editor September 7, 1995. 1991 Mathematics Subject Classification. Primary 65L05; Secondary 34C30. Key words and phrases. Isospectral flows, Runge-Kutta methods, conservation laws, unitary flows, Toda lattice equations. c 1997 American Mathematical Society 1461 License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 1462 MARI PAZ CALVO, ARIEH ISERLES, AND ANTONELLA ZANNA flows studied by Kac and von Moerbeke (cf. [KvM], [T]). We will return time and again to these flows in the sequel. It is important to point out that the aforementioned flows are obtained for very special choices of the matrix B(L). In the most general case the dynamics of (1) is still unknown or not yet fully understood. The most important qualitative feature of (1) is the isospectrality of the solution L(t). In other words the eigenvalues of the matrix L(t) are independent of time. This has been shown by Flaschka for the Toda lattice equations (see [F], [T]) but with greater generality his proof applies to all the flows that can be written in the form (1). Therefore it is often of essence to require that a numerical method for the initial value problem (1) retains isospectrality. So far, Moore–Mahony–Helmke have proposed in [MMH] an algorithm which produces an isospectral solution. Their algorithm is aimed to evaluate the eigenvalues of L0 rather than to approximate the solution of (1) with any degree of precision, and it is applicable just to the double-bracket flows. Instead we propose a considerably more general approach which allows us to produce an isospectral solution for the initial value problem (1) with an arbitrarily high order of accuracy. This new class of methods, which we call modified Gauss–Legendre Runge–Kutta (MGLRK) schemes, is based on a rendition of Flaschka’s theoretical approach in a computational form. This different technique is strongly motivated by the failure of standard ODE methods for the problem in hand. Isospectrality of (1) can be interpreted in the fol- lowing way. The solution L(t) lies on an intersection of several manifolds, each one corresponding to an integral for (1) that can be expressed in terms of a conservation law. We show that the most likely candidates, Runge–Kutta (RK) methods which retain quadratic conservation laws, fail since they cannot recover cubic integrals. This paper is organized as follows. Section 1 introduces some basic concepts for the problem in hand and describes the most ubiquitous isospectral flows. Section 2 is concerned with standard methods for ODEs. We derive the conditions that the coefficients of the numerical method have to obey in order to recover conservation laws. In particular, we prove that for Runge–Kutta schemes, conservation of qua- dratic and cubic integrals are conflicting requirements, thereby concluding that for d 3 no RK method can be isospectral. In Section 3 we introduce the modified Gauss–Legendre≥ RK methods and, finally, Section 4 is concerned with numerical examples. 1.2. The QR flow and the Toda flow. Given a function f which is analytic on the spectrum σ(L0)= λ1,λ2;::: ,λd of the matrix L0, we refer to (1) as a QR flow when { } (2) B(L)=f(L)+ f(L) : − − The subscripts ‘ ’ denote the (strictly) upper and lower triangular part of the matrix f(L) respectively.± The name QR flow (cf. [Sy], [Na1], [Na2], [DNT]) is due to the fact that for symmetric and positive definite L0 and f(x) = log x at integer time-steps the flow produces exactly the iterations of the familiar QR method for finding eigenvalues. Reversing the order in (2) is equivalent to reversing integration in time. Moser in [Mo] has shown that for t the flow L(t) tends to a diagonal matrix whose elements are the eigenvalues→±∞ of L(t), while Deift, Nanda and Tomei in [DNT] have shown that the convergence to the asymptotically stable equilibrium point is exponential. For t + , the eigenvalues of the flow (2) are arranged → ∞ License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use NUMERICAL SOLUTION OF ISOSPECTRAL FLOWS 1463 from the largest to the smallest, the other way around for reverse time integration. The flow retains the bandwidth of the initial matrix L0. This is clear for the Toda flow (see [T]) but, by virtue of the analyticity of f on σ(L0) it applies with greater generality also to (2) (cf. [DNT]). For f(x)=x, the identity function, and for tridiagonal L0, we obtain the Toda flow in a notation originally due to Lax (cf.[T]), inwhichcasewedenotethematrixL(t) in the form β1 α1 0 ::: 0 .. α1 β2 α2 . . (3) L(t)= 0 α .. .. 0 : 2 . .. .. .. . αd 1 − 0 ::: 0 αd 1 βd − d d × Occasionally we refer directly to the differential equations for the Toda flow (cf. [T]), namely 2 2 βk0 =2(αk αk1); (4) − − k =1;::: ;d; α0 =α(β β ); k k k+1 − k where here, as well as in the remaining part of this paper, we use the convention that α ,β = 0 whenever k 1; 2;::: ;d 1 and k 1;::: ;d respectively. k k 6∈ { − } 6∈ { } 1.3. The double-bracket flow. We refer to a double-bracket flow when (5) L0 =[L; [L; N]]; where N is a given d d symmetric matrix. Without loss of generality and observing that [L; [L; N]] = [[N;L× ];L], the flow can be written in the form (1), where B(L)=[N;L]=NL LN: − The flow was first introduced by Brockett in [B1], where the author shows that it can be formulated as a gradient flow evolving in a Riemannian manifold. In the same paper he has also proved that, for diagonal N and an initial matrix L0,both of them with distinct eigenvalues, the matrix function L(t) tends exponentially to a diagonal matrix as t + and the eigenvalues are then sorted accordingly to → ∞ the diagonal entries of N.IfL0or N have multiple eigenvalues, exponential (but not asymptotical) convergence is lost. He also showed how this flow can be used to diagonalize matrices, sort lists and solve linear programming problems. This flow in general does not retain the bandwidth of the initial matrix L0 except in the case N = κI diag 1; 2;::: ;d : ± { } In particular, when N = diag 1; 2;::: ;d , the double-bracket flow (5) is a refor- mulation of the Toda lattice. { } When N is nondiagonal, the analysis of convergence is essentially the same (cf. [B1]). To verify this, observe that N is symmetric, therefore it can be diagonalized by means of an orthogonal transformation. In other words, there exist an orthogonal matrix Q and a diagonal matrix Λ such that N = QΛQT ;QQT=QTQ=I: Next observe that, if Q is orthogonal and A; B are arbitrary d d matrices, it is true that × [QAQT ;QBQT]=Q[A; B]QT ; License or copyright restrictions may apply to redistribution; see https://www.ams.org/journal-terms-of-use 1464 MARI PAZ CALVO, ARIEH ISERLES, AND ANTONELLA ZANNA from which, letting L˜ = QT LQ, we deduce that the problem (5) is equivalent to L˜0 =[L;˜ [L;˜ Λ]]; with diagonal Λ.